Latent Inter-User Difference Modeling for LLM Personalization

Yilun Qiu¹, Tianhao Shi², Xiaoyan Zhao³, Fengbin Zhu¹, Yang Zhang¹, Fuli Feng²,

¹National University of Singapore, ²University of Science and Technology of China, ³The Chinese University of Hong Kong

This is the implementation of the Difference-aware Embedding-based Personalization (DEP) method proposed in our paper accepted by EMNLP 2025 Main Conference as an ORAL presentation.

📋 Catalogue

• 📋 Catalogue
• ⚙️ Environment Setup
• 📚 Dataset Preprocess
• ⌛️ Quick Start
  • Evaluation
  • Training
• 📊 Experimental Results
• 📖 Citation

⚙️ Environment Setup

bash install.sh

Feel free to take a look at install.sh to see what you need to run our project.

📚 Dataset Preprocess

The dataset we used in DEP is adapted from Amazon Reviews'23 and DPL.

bash run-create.sh

⌛️ Quick Start

Evaluation

To evaluate the performance of DEP, simply run the following command:

bash run-eval.sh

The evaluation script will download the model automatically from the Huggingface and test it.

Training

To train the model and select best-performing model, simply run the following commands:

bash run-train.sh
bash run-select.sh

It takes around 3 hours to train the model in a single H100.

📊 Experimental Results

📖 Citation

If you find our work useful, please kindly cite our paper:

@article{qiu2025latent,
  title={Latent Inter-User Difference Modeling for LLM Personalization},
  author={Qiu, Yilun and Shi, Tianhao and Zhao, Xiaoyan and Zhu, Fengbin and Zhang, Yang and Feng, Fuli},
  journal={arXiv preprint arXiv:2507.20849},
  year={2025}
}